Predicting 30-day Hospital Readmission with Publicly Available Administrative Database. A Conditional Logistic...

Area of Science:

• Healthcare Analytics
• Big Data in Medicine
• Predictive Modeling

Background:

• Hospital readmissions increase healthcare costs and patient distress.
• Current logistic regression models have limited prediction power for readmissions.
• Decision trees and random forests are complex for practitioners.

Purpose of the Study:

• To explore conditional logistic regression (CLR) for enhanced prediction accuracy of hospital readmissions.
• To compare CLR performance against standard classification models.

Main Methods:

• Analysis of HCUP statewide inpatient discharge data for heart failure patients.
• Application of standard logistic regression and decision trees to identify key variables.
• Stratification of data and application of logistic regression on strata, exploring interactions.
• Under-sampling for data imbalance and cross-validation for performance assessment.

Main Results:

• Conditional logistic regression models outperformed standard models, improving accuracy by nearly 20%.
• CLR achieved over 10% better sensitivity, potentially identifying 400-500 additional heart failure readmissions annually in California.
• Key predictors identified include discharge disposition, number of chronic conditions, and acute procedures.

Conclusions:

• Decision tree rules can guide cohort stratification for logistic regression.
• Exploring pairwise interactions in logistic regression models for different data strata is beneficial.
• Ad-hoc CLR models offer insights for predicting readmissions and developing post-discharge care strategies.